Halda Search Modeling for student search: 5 takeaways

In our October webinar, James Hamilton, Halda's vice president of data science, introduced Halda Search Modeling and showed how it decides where your next name buy and counselor trip should go. The full session is above. Here are five things to take back to your team.

1. Your inquiry map only shows where you've already looked

Many schools build their name buy from the places inquiries already come from, then add a few metros or states that feel right. James called it a reasonable method with one big blind spot. Inquiries reflect where you've marketed and where you pull organically, so buying to match them keeps you in the same places year after year. You never test the ZIP code three counties over, and you keep paying for neighboring ZIPs that don't convert.

Keep fishing where you already catch fish. Just don't let that be the whole plan. Every buy is a set of bets on regions, vendors, score and GPA bands, majors and volume. A search strategy tells you which bets to place, and it gives you a way to tell a weak region from a weak vendor or a weak student profile when names go quiet. If you're already pricing each recruitment line by cost per enrolled student, name buys are a smart place to start.

2. Start with one national model, then let thin markets lean on it

A common search model splits the country into markets first, then fits a separate model to each one. That works for a market with thousands of students. A market with 40 students gets a model built on 40 students, and in thin markets a ZIP with any history may hold one student. James put it plainly: a rate built on one student is a coin flip. The ranked ZIP list looks just as confident either way, and you rarely see how many students sit behind each market.

Halda Search Modeling starts with one national model trained on all of the students in your CRM. It learns which ZIPs sent you inquiries, applicants and enrolled students, then layers in two kinds of outside data:

• Census household data on income, poverty, housing, education and employment, limited to households raising children so retirees and empty nesters don't skew the picture
• High school data, including free and reduced lunch rates, per-pupil spending, the share of seniors who file a FAFSA and access to AP and calculus

Because the model learns what a ZIP that sends you students looks like, it can score every ZIP in the country, including the ones you've never touched. Each market then gets its own model, weighted by how much data it has. James compared it to a batting average: one strikeout in your first at-bat doesn't make you the worst hitter in the league. A market with 20 students leans mostly on the national model. A market with a couple thousand stands on its own.

3. Draw your markets, then test the map

The usual market map is a set of rings: your home state, the states around it, the next ring out and the rest of the country. Halda Search Modeling lets you draw that map yourself, let the data draw it or blend the two with a slider. Overrides stay put, so if one county belongs in market two, it stays there no matter where the slider sits.

You can build a scenario for each version of the map and compare how each one performs on results the model never trained on. In the live demo, a 50% blend outperformed the fully data-driven map, and the tool's recommended setup was two markets at a 40% blend. The recommendation comes from testing 63 combinations of six geographic tiers across the full range of blend points. You can change the map mid-cycle when applications come in, so it's never a one-time decision.

4. Judge a model on a class it never saw

A lot of search models get built once and reused, and few are ever checked against a class they haven't seen. Halda Search Modeling trains on earlier classes and tests on a later one. With 2024, 2025 and 2026 data, for example, it trains on 2024 and 2025 and scores 2026. James explained why that matters: holding out random students from every year tends to make a model look better than it is, because the future never looks exactly like the past.

Models also compete. Several model types go head to head on that holdout year in every market, and the best performer wins that market. Each market shows its holdout results, so you can see how well the model predicted before you trust it with a budget.

The target matters too. A model built on applicants alone treats a ZIP with 1,000 applicants and zero enrolled the same as one with 1,000 applicants and 200 enrolled. Halda Search Modeling counts applications and adds weight for enrollments, so it points at students who show up. ZIPs where purchased names go silent get a capped penalty that only lowers a score. Midway through the cycle, you can fold in this year's applications to date and retrain.

5. Use the same scores for vendors, travel and majors

The ranked ZIP list is the core of Halda Search Modeling, and the same model feeds six more tools:

• Vendor analysis shows how many names from each vendor inquired, applied and enrolled, market by market.
• Purchase criteria recommends a test score band, GPA band, vendor and majors for each ZIP, each with a confidence measure.
• High school analysis ranks every high school in the country, including schools that have never sent you an inquiry, so you can build a counselor travel plan.
• Major analysis shows which programs perform in which markets and maps your major names to the codes your list vendors use.
• Competitive landscape shows where the students you lose enroll and how your net price compares.
• Special population analysis tracks share and volume for the students your mission centers on. It never changes a ZIP score.

If you oversee more than one campus with separate applicant pools, each campus can get its own model.

-----------------

Watch the full webinar recording above to see James build markets live, compare scenarios and walk through the holdout results behind a score. Want to see it on your own data? Book a custom demo and we'll show you where your next class is most likely to come from.